CNN CNN METHOD AND DEVICE FOR TRANSFORMING CNN LAYERS TO OPTIMIZE CNN PARAMETER QUANTIZATION TO BE USED FOR MOBILE DEVICES OR COMPACT NETWORKS WITH HIGH PRECISION VIA HARDWARE OPTIMIZATION

摘要

The present invention provides a method for transforming a convolutional layer of a CNN including m convolutional blocks, (a) when an input image to be used by a computing device to determine a scaling parameter is obtained, (i) included in a kth convolutional block At least one kth initial weight of the kth initial convolutional layer, (ii) if (ii-1) k is 1, the input image, (ii-2) when k is from 2 to m, k(k−) 1) If the (k-1) feature map and (iii) (iii-1) k are 1, corresponding to the input image, output from the convolution block, corresponding to each channel included in the input image Each of the k-th scaling parameters and (iii-2) k is from 2 to m, one or more of the k-th scaling parameters corresponding to each of the channels included in the (k-1) characteristic map, Generating a kth quantization loss value-k is an integer from 1 to m-; (b) each of the kth optimal scaling parameters corresponding to each of the channels included in the (k-1)th feature map among the kth scaling parameters with reference to the kth quantization loss value by the computing device; Determining; (c) the computing device generating a kth scaling layer and a kth inverse scaling layer with reference to the kth optimal scaling parameter; (d) the computing device converts the kth initial convolutional layer to a kth convolutional convolutional layer using (i) kth scaling layer when k is 1, and (ii) k is 2 In the case of an integer up to m, using the kth scaling layer and the (k-1) inverse scaling layer, converting the kth initial convolution layer to the kth integrated convolution layer; It relates to a method and apparatus characterized by.